🔍 Read the full analysis: How OpenAI's Robotics Shift Connects To Runway's Praxis-1 on ThorstenMeyerAI.com
Prime made for students and young adults
- Fast, free delivery for dorm and study essentials
- Prime Video and Amazon Music included
- Member-only deals
TL;DR
OpenAI has reorganized its robotics teams, saying advances in general-purpose models have made it possible to apply those models directly to robot control. Runway has released Praxis-1, a dedicated open-weight robot model. Early results suggest general models can perform well across simulated tasks, while specialist models respond faster; the tests do not settle which approach will work best in deployed robots.
OpenAI has disbanded its stand-alone robotics team and moved its researchers into other parts of the company, Fortune reported November 13, while Runway released Praxis-1, a dedicated robot model, two days earlier. OpenAI told Fortune the change reflects progress in applying general-purpose models to robot control, not a retreat from robotics; early benchmark results show a trade-off between broad task performance and the speed needed for physical action.
Fortune reported that OpenAI confirmed the reorganization after the outlet found the company’s registered booth empty at the CoRL robotics conference in Austin. The robotics researchers now report to Chief Research Officer Mark Chen, while the applied robotics and hardware teams have moved to the Applied organization under President Greg Brockman. OpenAI had not published its own announcement as of November 14.
According to Fortune, an OpenAI spokesperson said its general-purpose frontier models, including GPT-6 Astra, had shown they could be applied directly to robotics. The spokesperson described robotics as part of the company’s long-term vision and said OpenAI remained committed to bringing general-purpose robots into the world. The company has said it is developing a humanoid robot; CEO Sam Altman has targeted a public demonstration in 2027.
Runway’s Praxis-1 is a 27-billion-parameter world action model released with open weights and code under the Apache 2.0 license, according to the supplied source material. Runway reports a 46.2% success rate on the RoboLab-120 benchmark, placing Praxis-1 fifth overall and ahead of other world-action and open-source robot models on that leaderboard. The reported result is not evidence that it will perform at the same level across different robots or real-world settings.
Robot control loop · explainer
Who decides the robot’s next move?
Every robot repeats one loop: cameras see, a model decides, the arm moves. The two bets in this story put a different kind of model into the decide step.
Option 1
A general model in the loop
Example: GPT-6 Astra in NVIDIA’s simulation study
18.56 s per model call on average in the direct set-up. Inference was 97.0% of the estimated episode time; the simulation paused while the model was thinking.
Option 2
A specialist robot model in the loop
Example: Runway’s Praxis-1, a world action model with open weights
About 0.55–0.6 s per call on one H100 server, and each call returns 1.6 s of motion. Needs fine-tuning data for each robot.
General model plans the next stepslow, outside the loop
Option 3 · illustrative
Both, in two layers
The idea Runway and the NVIDIA researchers both describe
Runway calls the two approaches “complementary”: an agent plans, a fast policy carries out each step. NVIDIA’s authors say the best way to combine them “remains an open question”.
Sources: NVIDIA Seattle Robotics Lab, “Astranomy” study of GPT-6 Astra, Oct 14, 2026 (simulation, direct set-up); Runway, Praxis-1 technical report, Nov 2026 (one H100 server). Loops and timelines are schematic; the two figures come from different set-ups, not from a head-to-head test.
Graphic: Powered by Thorsten Meyer AI
The Speed Gap in Robot Control
The reorganization reflects a strategic question for robotics: can general-purpose AI models become robot controllers, or do robots need models built specifically for fast physical action? A model that can reason across many tasks could reduce the need to train a separate policy for each one. But a robot must also respond on a timescale that fits the movement and environment, where slow model calls can limit practical use.
In an NVIDIA researchers’ simulation study, GPT-6 Astra reached 74.8% success on RoboLab-120 in an agentic setup, compared with 51.0% for the strongest trained robot policy reported on the leaderboard. The study’s direct setup averaged 18.56 seconds per model call, however, and the simulation paused while the model generated an answer. Runway reports that Praxis-1 takes about 0.55 to 0.6 seconds per call for 1.6 seconds of motion. These figures come from different setups and should not be treated as a direct, independent comparison.
Runway also says that, in its own test on one of its robots, GPT-6 Astra took 15.6 times longer than Praxis-1 to complete three tasks and cost 51 times more per completed task. That is a company-reported test, not a broad evaluation. Together, the available results point to possible roles for both approaches: general models may help with flexible reasoning, while faster specialist models handle immediate control. Runway itself calls the approaches complementary.
Popular right now
KitchenAid Artisan 5-Qt Tilt-Head Stand Mixer, KSM150PS, Almond Cream
Roborock Q7 M5+ Robot Vacuum and Mop
De'Longhi Classic Signature Espresso Machine with Adjustable Milk Frother
As an affiliate, we earn on qualifying purchases.
OpenAI’s change follows earlier shifts in its robotics work. The company wound down an earlier stand-alone robotics team in 2020, then rebuilt a team in 2024 and partnered with humanoid-robot maker Figure. Figure said it would develop its own robot AI when that partnership ended in February 2025. Fortune described the current change as the second time OpenAI has closed a stand-alone robotics team.
At the start of 2026, OpenAI researchers expected to build on video models such as Sora, according to the spokesperson quoted by Fortune. The company now says its newer general-purpose models can be applied directly to robotics. This is the explanation OpenAI gave for the restructuring. Fortune also reported that OpenAI faces pressure to control costs and focus on core products, but the supplied reporting does not establish that cost pressure caused this particular decision.
The NVIDIA study cited in the reports was published October 14, 2026, as a CoRL 2026 late-breaking result. It tested GPT-6 Astra on RoboLab-120, a simulation of 120 tabletop tasks for a single robot arm. In addition to the 74.8% agentic result, the direct-policy setup scored 66.3%, while an orchestrated setup, in which Astra reviewed proposals from a trained robot model, scored 66.5%. These are simulation results, not demonstrations of sustained performance in homes, factories or other real-world environments.
“The two approaches are complementary.”
— Runway
What the Tests Do Not Show
The available figures do not establish which model approach is best for real-world robot deployment. RoboLab-120 is a simulation benchmark, and the supplied material does not provide comparable independent tests across multiple robot bodies, environments or long-running tasks. The benchmark scores also measure task success, while latency and cost measure different aspects of performance.
Runway’s timing and cost comparison comes from a test it conducted on one of its own robots and covered three tasks. The provided material does not detail enough about the test conditions to establish how those numbers would generalize. OpenAI has not released a public statement explaining the reorganization, and the company has not published the full rationale or a revised robotics roadmap. It is also unclear whether OpenAI plans to continue developing its own humanoid hardware on the previously stated schedule.
The Next Tests for Both Models
The next useful evidence will be independent, comparable evaluations that test task success, response time and cost on the same robots and tasks. Results in physical environments would also help show whether simulated performance transfers to settings where objects, lighting and movement are less predictable.
OpenAI’s next public milestone to watch is the humanoid demonstration Altman has targeted for 2027, though the supplied reporting does not confirm a date or say whether the plan has changed. Runway has released Praxis-1’s weights and fine-tuning code under Apache 2.0, allowing developers to test and adapt the model. For now, the reorganization and release mark competing organizational choices—not a settled answer about how robots will be controlled.
Key Questions
What changed in OpenAI’s robotics organization?
According to Fortune, OpenAI disbanded its stand-alone robotics team. Its researchers moved under Chief Research Officer Mark Chen, and its applied robotics and hardware teams moved to the Applied organization under President Greg Brockman.
What is Praxis-1?
Praxis-1 is Runway’s 27-billion-parameter world action model for robotics. Runway released its weights and fine-tuning code under the Apache 2.0 license, according to the supplied source material.
Did GPT-6 Astra outperform Praxis-1?
There is no single direct result establishing that. GPT-6 Astra scored 74.8% in an NVIDIA researchers’ RoboLab-120 simulation setup, while Runway reports 46.2% for Praxis-1. The results come from different evaluations, and Runway’s robot test is separate from the NVIDIA benchmark.
Why might a specialist model still be useful?
Robot control often requires fast responses. The reported NVIDIA study averaged 18.56 seconds per call for GPT-6 Astra in its direct setup, while Runway reports about 0.55 to 0.6 seconds per Praxis-1 call for 1.6 seconds of motion. Those measurements use different test setups, but they illustrate why response speed matters.
Does the reorganization mean OpenAI is leaving robotics?
OpenAI’s spokesperson told Fortune that robotics remains part of the company’s long-term vision and said it remains committed to general-purpose robots. The company has not published its own announcement about the change, so details of its future robotics plans remain limited.
Source: ThorstenMeyerAI.com
